Magnetic resonance hypergraph contrastive learning-based autism diagnosis apparatus and method, and storage medium

Abstract

A magnetic resonance hypergraph contrastive learning-based autism diagnosis method, comprising a pre-training stage and a fine-tuning stage and used for using limited functional magnetic resonance data to identify autism spectrum disorder (ASD). The present invention uses three frequency bands of the amplitude of low-frequency fluctuations (ALFF) as node features, and uses Pearson correlation coefficients (PCC) of average functional magnetic resonance time series of different brain regions as edge weightings. The pre-training stage focuses on self-supervised learning, and combines hyper-edge perception contrastive learning and adversarial training (SSL-HCLA), so as to extract high-order relational features from each brain region of interest (ROI). In the fine-tuning process, pre-trained SSL-HCLA is refined, and a classification module is trained to accurately identify ASD.

Description

Autism diagnosis apparatus and method based on magnetic resonance hypergraph contrast learning and storage medium TECHNICAL FIELD

[0001] The present application relates to the field of medical rehabilitation, in particular to an autism diagnosis apparatus based on magnetic resonance hypergraph contrast learning and a storage medium. BACKGROUND

[0002] Autism spectrum disorder (ASD) is a developmental disorder characterized by severe challenges in social communication, interaction, and behavior. According to the recent data from the Centers for Disease Control and Prevention (CDC), approximately one in 36 children is diagnosed with autism. Despite the prevalence of ASD, current diagnosis mainly relies on traditional behavioral assessments. These methods are often subjective, leading to potential delays in identifying early symptoms and increased risk of misdiagnosis. Therefore, in psychiatry and neuroscience, more and more people are looking for objective biomarkers of ASD. This shift aims to facilitate early diagnosis and timely intervention, which is crucial for improving treatment outcomes.

[0003] Resting-state functional magnetic resonance imaging (rs-fMRI) is a technique that measures blood oxygen level dependent (BOLD) signals in the brain without requiring subjects to perform specific tasks. This technique has been widely used to identify potential neuroimaging biomarkers for various mental illnesses, and the exploration of functional connectivity networks (FCNs) pairs the relationships between regions of interest (ROIs) in each subject's brain. FCNs help us understand brain organization patterns and diagnose mental illnesses such as ASD, attention deficit / hyperactivity disorder (ADHD), depression, bipolar disorder, and anxiety. However, previous studies usually extract handcrafted network features (such as node degree and clustering coefficient) from FCNs, and then use these features in predictive models for ASD diagnosis. This approach limits the accuracy of diagnosis and is heavily dependent on expert knowledge. Recently, a large number of machine learning models have been applied to FCN analysis for automatic diagnosis of brain diseases, such as support vector machines (SVMs) for differentiating multiple system atrophy and Parkinson's disease, logistic regression for early detection of Alzheimer's disease, and deep neural networks (DNNs) for identifying schizophrenia. However, these models often vectorize FCN matrices, thereby ignoring important topological structure information between ROIs. To reveal discriminative patterns in brain networks, FCNs should be treated as irregular graph-structured data and analyzed as a whole, where each ROI represents a vertex and its associated FC profile serves as a vertex feature.

[0004] To this end, brain disease recognition can be seen as a graph classification problem, where nodes represent different brain regions and edges represent the strength of connections between these regions. The main task of graph classification problem is to learn the representation of graph-structured data. Unlike shallow models, deep neural networks can learn a hierarchical representation of functional brain networks, enabling them to capture complex interactions between brain regions. Ktena et al. proposed a Siamese GCN model that learns a graph similarity measure between fcn to distinguish between autism patients and healthy controls. Similarly, Qin et al. studied a graph embedding learning (GEL) model for diagnosing major depressive disorder. The model first learns FCN embeddings using GCN, then applies a fully connected layer activated by a softmax function for final classification. However, functional connectivity networks (FCNs) constructed using traditional graph-based methods, such as Pearson correlation and partial correlation using graphical Lasso, are limited to capturing relationships between pairs of regions of interest (ROIs), i.e., second-order relationships. This approach ignores interactions between multiple ROIs, or high-order relationships (Figure 1, let V = {v1, v2, v3, v4, v5, v6, v7, v8, v9, v 10} be a set of regions of interest (e.g., vertices); Figure 1 (a): FCN represented as a graph, where each edge is represented by a black line segment, and each edge connects two regions of interest; Figure 1 (b): FCN represented as a hypergraph, e1 = {v7, v8, v9}, e2 = {v1, v2, v6, v7, v8, v9}, e3 = {v3, v4, v5, v6, v9}, e4 = {v4, v5, v9, v 10}).

[0005] Given the complexity of brain networks, high-order relationships can be crucial to understanding neural processes related to cognition and behavior. Moreover, previous studies have not demonstrated significant improvements in brain disease diagnosis. We hypothesize that this performance gap is mainly due to two main limitations of deep neural networks: (1) lack of sufficient training data (2) lack of supervision.

[0006] Self-supervised learning (SSL) has gained attention as a promising approach that leverages the inherent structure of data to provide supervision, making it particularly valuable in cases where labeled data is limited. While SSL has been successful in computer vision (CV) and natural language processing (NLP), it has been rarely used in brain network analysis, where self-supervised learning is more important to address the problems of limited data and lack of supervision. Therefore, it is crucial to consider the topological structure of brain networks when developing graph self-supervised learning. Two key aspects need to be addressed:

[0007] (1) How to develop a self-supervised encoder model that captures richer representations by exploiting the graph structure of brain networks?

[0008] (2) How to use these captured features to improve the accuracy of autism diagnosis?

[0009] Application content

[0010] The purpose of the present application is to provide an autism diagnosis device based on magnetic resonance supergraph contrast learning and a storage medium to solve the technical problems described in the background art.

[0011] An autism diagnosis device based on magnetic resonance supergraph contrast learning, comprising: a self-supervised learning module, a graph classification model (fine-tuning stage); self-supervised learning includes super-edge perception contrast learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes ASD and HC (Healthy Control) subjects by analyzing brain connectivity patterns and functional relationships.

[0012] Compared with related technologies, the advantages and positive effects of the present application are that SSL-HCLA includes creating an original supergraph and an edge-reduced version supergraph for contrast learning and feature extraction. In this process, the model distinguishes between the two supergraph versions by deleting edges using a trainable Bernoulli mask, which promotes the learning of invariant features. The filtered features are robust to supergraph changes and play a crucial role in disease diagnosis. After training, the remaining edge weights can be analyzed as potential biomarkers. The graph classification model (in the fine-tuning stage) uses a linear classifier to classify these extracted features, and the classifier distinguishes ASD and HC subjects by analyzing brain connectivity patterns and functional relationships. The proposed model captures complex multi-node interactions through supergraph representation and generates meaningful representations under limited labeled data using self-supervised learning, thereby improving diagnostic accuracy and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0013] These and other features and advantages will become more apparent from the following detailed description of illustrative embodiments thereof, taken in conjunction with the accompanying drawings and claims.

[0014] BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a schematic diagram of a functional connectivity network constructed by a conventional graph-based method.

[0016] Figure 2 is a model structure diagram of an autism diagnosis device based on magnetic resonance supergraph contrast learning.

[0017] Fig. 3 is a distribution map of an embodiment of an autism diagnosis device and storage medium based on magnetic resonance hypergraph contrastive learning and its connection on the cortical surface.

[0018] Fig. 4 is a schematic diagram of an embodiment of an autism diagnosis device and storage medium based on magnetic resonance hypergraph contrastive learning and its connection strength increase. DETAILED DESCRIPTION

[0019] It should be understood that, although illustrative implementations of one or more embodiments are provided below, the disclosed systems and / or methods can be implemented using any number of techniques, whether currently known or in existence. The disclosure should in no way be limited to the illustrative implementations, examples, drawings, and techniques in the following detailed description, and should be given the widest scope consistent with the claims and their equivalents.

[0020] As shown in the accompanying drawings 1-4 of the specification:

[0021] An autism diagnosis device based on magnetic resonance hypergraph contrastive learning includes a self-supervised learning module, a graph classification model (fine-tuning stage); self-supervised learning includes hyperedge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes ASD and HC subjects by analyzing brain connectivity patterns and functional relationships.

[0022] In this embodiment, SSL-HCLA includes creating an original hypergraph and an edge-reduced version of the hypergraph for contrastive learning and feature extraction. In this process, the model distinguishes between the two hypergraph versions by deleting edges using a trainable Bernoulli mask, which promotes the learning of invariant features. The filtered features are robust to hypergraph changes and play a crucial role in disease diagnosis. After training, the remaining edge weights can be analyzed as potential biomarkers. The graph classification model (in the fine-tuning stage) uses a linear classifier to classify these extracted features, and the classifier distinguishes ASD and HC subjects by analyzing brain connectivity patterns and functional relationships. The proposed model captures complex multi-node interactions through hypergraph representation and generates meaningful representations under limited labeled data using self-supervised learning, thereby improving diagnostic accuracy and robustness.

[0023] Specifically, an autism diagnosis device based on magnetic resonance hypergraph contrastive learning further includes a data preprocessing module, which includes several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice time, and correcting sensitivity distortion.

[0024] Specifically, the data preprocessing module involves using inter-frame displacement, global signal, and average tissue signal as covariates for removing confounding factors, and then the fMRI volume is registered to the standard MNI152 space and is regressed.

[0025] Specifically, the autism diagnosis device based on magnetic resonance hypergraph contrast learning further comprises: using an AAL1 atlas to divide the brain into 116 regions. The average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels in the region; then, the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the Pearson correlation coefficient (PCC) between the pair of average time series is used to calculate the full functional connectivity (FC) matrix. The brain of each subject is represented as a hypergraph using the AAL116 atlas.

[0026]

[0027] Specifically, the autism diagnosis device based on magnetic resonance hypergraph contrast learning further comprises constructing a brain hypergraph:

[0028] The brain hypergraph is constructed from an A116 atlas, which divides the fMRI image into a plurality of regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes an edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated according to the Pearson correlation coefficient between the average time series of two different regions.

[0029] The node features are derived from the amplitudes of low-frequency fluctuations (ALFFs) in three frequency bands of the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and typical frequency band (0.01-0.08 Hz). ALFFs are defined as the total power in these low-frequency ranges, obtained by Fourier transform of the average time series. Some previous studies have demonstrated the effectiveness of using ALFFs as node features. The hypergraph constructed in this way is represented as , where symbolizes the set of nodes, represents the set of hyperedges, is the diagonal matrix of edge weights, ; the hypergraph is represented by the incidence matrix , and the rules for constructing this incidence matrix are described as follows:

[0030]

[0031] where the degree of a node , the degree of an edge is defined as , a hyperedge is denoted by . The diagonal matrices representing the node degrees and hypergraph degrees are denoted by and , respectively. In , each element specifies the number of hyperedges linked to the corresponding node, while in , each element represents the number of nodes associated with the corresponding hyperedge. With the incidence matrix of the hypergraph, the diagonal matrices and can be derived to encapsulate the nodes and hyperedges. Given the node features , the node embeddings are then updated by the process of hyperedge convolution.

[0032]

[0033] where denotes the node features of the -th layer, is a learnable parameter, denotes a nonlinear activation function.

[0034] In particular, the autism diagnosis device based on magnetic resonance hypergraph contrast learning further includes a masked hypergraph encoder:

[0035] The masked hypergraph encoder is used for node embedding learning and edge embedding: let the fMRI data construct a hypergraph . This hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature connection, and an MLP layer. Each layer updates the node features through a message passing process. Let the set of neighboring nodes of a given node be denoted by . The message passing process is then defined as follows:

[0036]

[0037] where denotes a hyperedge convolution block that converts the neighboring node features and edge weights into an aggregated vector. The function is an MLP layer, which is a trainable function that maps the current node representation and the aggregated vector to a new representation. The message passing process is as follows:

[0038]

[0039] The above process can be written in matrix form:

[0040]

[0041] in express Normalized node features, matrix Include The corresponding row, and , , , These are trainable parameters. , .symbol This represents the batch normalization operation. The hypergraph module aims to learn the latent representations of node features. Then connect them to form embedded edge features. .

[0042] Specifically, an autism diagnostic device based on magnetic resonance hypergraph contrastive learning also includes a masked hypergraph autoencoder:

[0043] Edge features are input into the MLP layer to generate parameters for a Bernoulli distribution, and then randomly discarded edges during sampling. The MLP layer includes a layer with trainable weights. and bias A linear layer, followed by a Activation function and another with weights and bias The linear layers ultimately form a sigmoid function that maps the output to the range (0,1). This process is mathematically defined as:

[0044]

[0045] In this way, the marginal features are transformed into scalars representing the parameters of the Bernoulli distribution.

[0046] Given parameters Discard the indicator at the edge of each edge. Perform sampling, thus ,in This indicates that edges are discarded. Matrix This forms a binary (0 or 1) Bernoulli mask B. To ensure gradient backpropagation, the sampling process needs to be reparameterized. The reparameterization technique used by Luo et al. is for this purpose.

[0047]

[0048] where , is a temperature parameter governing the smoothness of the reparameterized sampling function. If falls in the interval , the expression inside the sigmoid function is positive, if it lies in the interval , it is negative. Thus, the output converges to 1 with probability . As approaches 0, becomes more and more similar to a binary indicator of the sample.

[0049] The Bernoulli mask is applied to the affinity matrix by element-wise multiplication, masking specific edges when . Thus, the input hypergraph is converted into an edge-deleted hypergraph, preserving the same number of nodes. The process of generating the parameter matrix from the hypergraph g = (V, E, W) is denoted by e, and the whole data augmentation process can be denoted as:

[0050]

[0051] In particular, the pre-training phase of self-supervised adversarial contrastive learning, the input hypergraph and its augmented version are processed by a hypergraph encoder and a projection head to extract hypergraph features, these inputs pass through a weight-shared edge encoder and node encoder block.

[0052] To train the feature extractor and the parameters in the hypergraph augmentation process, a loss function is used, which forces two feature vectors to approach if they come from the same hypergraph, and to move away if they come from different hypergraphs. In this way, the trained feature extractor can retain the most important information while removing excessive information in the graph.

[0053] Since the training is done in a batched manner, such a loss is also applied in a batched manner. Specifically, the loss is defined by the infoMax principle, which aims to maximize:

[0054]

[0055] where denotes the set of hypergraphs in a batch, denotes its cardinality. The similarity measure, denoted by sim, is defined as the cosine of the angle between two input vectors:

[0056]

[0057] maximizing This can be done by keeping all edges. To encourage more edges to be removed, a regularization term aims to encourage more edges to be removed and minimize all average of values

[0058]

[0059] Finally, the overall objective function is defined as:

[0060]

[0061] The optimal values of and are determined by gradient descent / ascent on the respective parameters.

[0062] In particular, the fine-tuning stage of the hypergraph classification, in the fine-tuning stage, the pre-trained SSL-CLAH model is further optimized to enhance the feature representation, especially to identify important and abnormal regions related to ASD. Then the refined features from the SSL-CLAH block are input into the classification block. The final output is a probability-based decision indicating the likelihood of the subject having autism spectrum disorder (ASD) or healthy control (HC). This decision is intuitively represented in the form of probability, thus helping the diagnosis process. The loss function used in the fine-tuning stage is the cross-entropy loss, defined as:

[0063]

[0064] where denotes the label of the pth graph, denotes the prediction for the pth subject, is the number of samples.

[0065] Another embodiment, a method for autism diagnosis based on magnetic resonance hypergraph contrastive learning, comprising: a self-supervised learning stage, a graph classification model (fine-tuning stage); the self-supervised stage includes hyperedge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, the linear classifier classifies the extracted features, and the linear classifier distinguishes ASD and HC subjects by analyzing brain connectivity patterns and functional relationships.

[0066] In particular, a method for autism diagnosis based on magnetic resonance hypergraph contrastive learning, further comprising a data preprocessing stage, including several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice timing, and correcting sensitivity distortion.

[0067] Specifically, a method for autism diagnosis based on magnetic resonance hypergraph contrast learning, in the data preprocessing stage, mixed factors are removed by using inter-frame displacement, global signal and average tissue signal as covariates, after the fMRI volume is registered to the standard MNI152 space, it is regressed.

[0068] Specifically, the method for autism diagnosis based on magnetic resonance hypergraph contrast learning further comprises: using an AAL1 atlas to divide the brain into 116 regions; the average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels in the region; then, the amplitude of the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the Pearson correlation coefficient (PCC) between the pair of average time series is used to calculate the full functional connectivity (FC) matrix; the brain of each subject is represented as a hypergraph by using an AAL116 atlas.

[0069] Specifically, the method for autism diagnosis based on magnetic resonance hypergraph contrast learning further comprises constructing a brain hypergraph:

[0070] The brain hypergraph is constructed by an A116 atlas, and the A116 atlas divides the fMRI image into a plurality of regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes an edge of the graph; the average time series in each ROI is calculated by averaging the BOLD signals of all voxels in the region; then, the edge weight is calculated according to the Pearson correlation coefficient between the average time series of two different regions;

[0071] The node feature is derived from the amplitude of the low-frequency fluctuation (ALFFs) in three frequency bands of the BOLD signal: Slow-5 (0.01-0.027Hz), Slow-4 (0.027-0.073Hz) and typical frequency band (0.01-0.08Hz); ALFFs is defined as the total power in these low-frequency ranges, which is obtained by Fourier transform of the average time series; some previous studies have proved the effectiveness of using ALFFs as node features; the hypergraph constructed therefrom is represented as , wherein symbolizes a set of nodes, symbolizes a set of hyperedges, symbolizes a diagonal matrix of edge weights, The hypergraph is represented by an affinity matrix The rules for constructing this affinity matrix are described as follows:

[0072]

[0073] where the degree of the node , define the degree of edge as , hyperedge is denoted as ; the diagonal matrices representing the node degree and hypergraph degree are denoted as and ; in , each element specifies the number of hyperedges linked to the corresponding node, while in , each element represents the number of nodes associated with the corresponding hyperedge; with the incidence matrix of the hypergraph, the diagonal matrices and can be derived to encapsulate the nodes and hyperedges; given the node features , the node embeddings are then updated through the process of hyperedge convolution;

[0074]

[0075] where denotes the node features of the layer, is the learnable parameter, and denotes the nonlinear activation function.

[0076] Specifically, the autism diagnosis method based on magnetic resonance hypergraph contrast learning also sets a mask hypergraph encoder:

[0077] The mask hypergraph encoder is used for node embedding learning and edge embedding: let the fMRI data construct a hypergraph ; this hypergraph undergoes a series of transformations, including a hyperedge convolution block, feature connection, and an MLP layer; each layer updates the node features through a message passing process; let the set of adjacent nodes of a given be denoted as ; the message passing process is then defined as follows:

[0078]

[0079] where denotes the hyperedge convolution block that converts the adjacent node features and edge weights into an aggregated vector; the function is an MLP layer, which is a trainable function that maps the current node representation and the aggregated vector to a new representation; the message passing process is as follows:

[0080]

[0081] The above process can be written in matrix form:

[0082]

[0083] in express Normalized node features, matrix Include The corresponding row, and , , , These are trainable parameters. , ;symbol This represents the batch normalization operation; the hypergraph module aims to learn the latent representations of node features. Then connect them to form embedded edge features. .

[0084] Specifically, an autism diagnostic method based on magnetic resonance hypergraph contrastive learning also includes a masked hypergraph autoencoder:

[0085] Edge features are input into the MLP layer to generate parameters for a Bernoulli distribution, and then randomly discarded edges during sampling. The MLP layer includes a layer with trainable weights. and bias A linear layer, followed by a Activation function and another with weights and bias The linear layer ultimately forms a sigmoid function that maps the output to the range (0,1); this process is mathematically defined as:

[0086]

[0087] In this way, the edge features are transformed into scalars, representing the parameters of the Bernoulli distribution;

[0088] Given parameters Discard the indicator at the edge of each edge. Perform sampling, thus ,in This indicates that edges are discarded; matrix A binary (0 or 1) Bernoulli mask B is formed; in order to ensure gradient backpropagation, the sampling process needs to be reparameterized; the reparameterization technique of Luo et al. is used for this purpose;

[0089]

[0090] where , is a temperature parameter governing the smoothness of the reparameterized sampling function; if falls in the interval , the expression inside the sigmoid function is positive, if it lies in the interval , it is negative; therefore, the output converges to 1 with probability ; as tends to 0, becomes more and more similar to a binary indicator of the sample;

[0091] The Bernoulli mask is applied to the affinity matrix by element-wise multiplication, masking a particular edge when ; therefore, the input hypergraph is converted into an edge-deleted hypergraph, preserving the same number of nodes; the process of generating the parameter matrix from the hypergraph g = (V, E, W) is denoted as e, and the whole data augmentation process can be represented as:

[0092] .

[0093] Specifically, the autism diagnosis method based on magnetic resonance hypergraph contrast learning, in the pre-training stage, the self-supervised adversarial contrast learning, the input hypergraph and its enhanced version are processed by the hypergraph encoder and the projection head to extract the hypergraph features, these inputs pass through a weight-shared edge encoder and node encoder block;

[0094] In order to train the feature extractor and the parameters in the hypergraph enhancement process, a loss function is used, which forces the two feature vectors to approach if they come from the same hypergraph, and to move away if they come from different hypergraphs.

[0095] Specifically, the autism diagnosis method based on magnetic resonance hypergraph contrast learning, the loss is defined by the infoMax principle, the goal is to maximize:

[0096]

[0097] where denotes the set of hypergraphs in a batch, denotes its cardinality; the similarity measure, denoted as sim, is defined as the cosine of the angle between two input vectors:

[0098]

[0099] maximizing This can be done by keeping all edges; to encourage more edges to be removed, a regularization term aims to encourage more edges to be removed and minimize the average of all the average of the values

[0100]

[0101] Finally, the total objective function is defined as:

[0102]

[0103] The optimal values of and are determined by gradient descent / ascent on the respective parameters.

[0104] Specifically, one kind of autism diagnosis method based on magnetic resonance hypergraph contrast learning, the hypergraph classification in the fine-tuning stage, in the fine-tuning stage, the pre-trained SSL-CLAH model is further optimized to enhance the feature representation, especially to identify important and abnormal regions related to ASD; then the refined features from the SSL-CLAH block are input into the classification block; the final output is a probability-based decision indicating the likelihood of the subject having autism spectrum disorder (ASD) or healthy control (HC); this decision is intuitively represented in the form of probability, thereby helping the diagnosis process; the loss function used in the fine-tuning stage is cross-entropy loss, defined as:

[0105]

[0106] where represents the label of the pth graph, represents the prediction for the pth subject, is the number of samples.

[0107] A storage medium for autism diagnosis based on magnetic resonance hypergraph contrast learning, which stores computer programs (instructions) thereon, the computer programs (instructions) are used to implement the autism diagnosis method based on magnetic resonance hypergraph contrast learning described above.

[0108] The autism diagnosis device, method and storage medium based on magnetic resonance hypergraph contrast learning related to the present application use the data sets from the public database Autism Brain Imaging Data Exchange I (ABIDE I) and ABIDE II, and the experimental results show that (from the results in Table 2), the ABIDE I and ABIDE II data sets prove the superior performance of the model proposed in terms of ASD classification. The technical solution related to the present application always reaches the highest indicators in all evaluation categories. For ABIDE I, the technical solution related to the present application reaches an accuracy of 93.2%, an AUC of 93.1%, a sensitivity of 92.9%, and a specificity of 95.1%, which is significantly better than other models such as ASD-DiagNet, GCN and AWSO-DBN; similarly, in the ABIDE II data set, the technical solution related to the present application maintains its superior performance, with an accuracy of 88.9%, an AUC of 88.8%, a sensitivity of 88.5%, and a specificity of 91.7%. These results clearly show that the technical solution related to the present application has substantially improved in terms of accuracy, AUC, sensitivity and specificity compared to other existing methods, proving its robustness and effectiveness in ASD classification.

[0109]

[0110] The autism diagnosis device, method and storage medium based on magnetic resonance hypergraph contrast learning related to the present application provide tools for personalized diagnosis of autism, assist clinical decision-making; extensive ablation studies confirm the robustness of the proposed model; in addition, the explanatory results emphasize the key functional connections and brain regions related to ASD.

[0111] A gradient-based saliency map method is used for model explanation to identify important brain regions and connections that contribute to classification. To analyze the edge features, the edge feature gradients of the ASD group are transposed and averaged to generate edge-symmetric saliency maps. Compared with HC, the edge features of ASD are increased with positive gradients and decreased with negative gradients. The top 66 connections (top 1%) with the highest absolute values are identified from a total of 6670 connections (0.5 x 116 x 115). These connections are shown in FIG. 3(a), and their distribution on the cortical surface is shown in FIG. 3(b).

[0112] As shown in Figure 3: Model interpretability for ABIDE I. Figure 3(a): All 66 selected functional connectivity features are shown in this circular plot; Figure 3(b) Functional links of the 66 brain regions, where darker and lighter colored lines represent increased and decreased functional connectivity strength in ASD vs. normal controls, respectively; Figure 3(c): Increased (left) and decreased (right) nodal features in ASD vs. healthy controls, where markers of different shapes represent Slow-5, Slow-4, and classical ALFF features, respectively. Abbreviations: R, right; L, left; MFG, middle frontal gyrus; ORBmid, middle orbital gyrus; ROL, Rolandic operculum; SFGmed, medial superior frontal gyrus; ORBsupmed, superior medial orbital gyrus; DCG, dorsal cingulate gyrus; ANG, angular gyrus; PCUN, precuneus; STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus.

[0113] As shown in Figure 3: Model interpretability for ABIDE I. Figure 3(a): All 66 selected functional connectivity features are shown in this circular plot; Figure 3(b) Functional links of the 66 brain regions, where darker and lighter colored lines represent increased and decreased functional connectivity strength in ASD vs. normal controls, respectively; Figure 3(c): Increased (left) and decreased (right) nodal features in ASD vs. healthy controls, where markers of different shapes represent Slow-5, Slow-4, and classical ALFF features, respectively. Abbreviations: R, right; L, left; MFG, middle frontal gyrus; ORBmid, middle orbital gyrus; ROL, Rolandic operculum; SFGmed, medial superior frontal gyrus; ORBsupmed, superior medial orbital gyrus; DCG, dorsal cingulate gyrus; ANG, angular gyrus; PCUN, precuneus; STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus.

[0114] As shown in Figure 4: Model interpretability for ABIDE II. Figure 4(a): All 66 selected functional connectivity features are shown in this circular plot; Figure 4(b) Functional links of the 66 brain regions, where darker and lighter colored lines represent increased and decreased functional connectivity strength in ASD vs. normal controls, respectively; Figure 4(c): Increased (left) and decreased (right) nodal features in ASD vs. healthy controls, where markers of different shapes represent Slow-5, Slow-4, and classical ALFF features, respectively. Abbreviations: R, right; L, left; MFG, middle frontal gyrus; ORBmid, middle orbital gyrus; ROL, Rolandic operculum; SFGmed, medial superior frontal gyrus; ORBsupmed, superior medial orbital gyrus; DCG, dorsal cingulate gyrus; ANG, angular gyrus; PCUN, precuneus; STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus.

[0115] The study also found that ASD patients had abnormal features of the cuneus / calcarine fissure and the dorsolateral prefrontal cortex (DLPFC) compared to the HC group. Meta-analysis showed that autistic individuals showed less prefrontal activity in perceptual tasks and greater extracortical supplementation in visual processing. The cuneus is associated with visual processing and retrieval of episodic memory. The DLPFC is connected to the orbitofrontal cortex, temporal lobe, and neocortex (including the posterior temporal, parietal, and occipital regions), and is involved in goal-directed thinking and action, and is associated with the non-social abnormalities of ASD, such as restrictive and repetitive behaviors, or reduced ideas of self-reference. Therefore, the DLPFC has been used as a target for transcranial electrical or magnetic stimulation to treat ASD. The cerebellum is also a key node for motor learning, language comprehension, social skills, visual-spatial performance, and memory functions, which overlap with the abnormal regions observed in ASD patients. It is noted that the cerebellar region has reduced structural or functional connectivity with the dorsolateral and the medial prefrontal cortex, which is associated with increased repetitive behaviors and social deficits in ASD mouse models and ASD patients.

[0116] Embodiments of the present application can be provided as a method, a system, or a computer program product.

[0117] In a possible implementation, an embodiment of the present application provides a computer readable storage medium storing program codes, which, when executed on a computer, cause the computer to perform the method embodiments described above.

[0118] In a possible implementation, an embodiment of the present application provides a computer program product, which, when executed on a computer, causes the computer to perform the method embodiments described above.

[0119] Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer usable program code.

[0120] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a specific manner, such that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device that implements the functions specified in one or more flows of the flowchart and / or one or more blocks of the block diagram.

[0121] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks.

[0122] Obviously, persons having ordinary skill in the art can make various modifications and variations to the embodiments of the present application without departing from the scope of the present application. Thus, if these modifications and variations of the embodiments of the present application belong to the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations. In the various embodiments of the present application, the terms and / or descriptions are consistent between the various embodiments if not specially stated and there is no logical conflict, and the terms and / or descriptions between the various embodiments can be referred to each other, and the technical features in different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0123]

Claims

1. An autism diagnosis apparatus based on magnetic resonance hypergraph contrastive learning, comprising: Self-supervised learning module, graph classification model (fine-tuning stage); self-supervised learning includes hyper-edge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); the graph classification model includes a linear classifier, the linear classifier classifies the extracted features, and the linear classifier distinguishes ASD and HC subjects by analyzing brain connection patterns and functional relationships.

2. The autism diagnosis device based on magnetic resonance supergraph contrastive learning according to claim 1, further comprising a data preprocessing module, including several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice timing, and correcting sensitivity distortion.

3. The autism diagnosis device based on magnetic resonance supergraph contrastive learning according to claim 2, wherein the data preprocessing module removes confounding factors by using inter-frame displacement, global signal and average tissue signal as covariates, and after registering the fMRI volume to the standard MNI152 space, it is regressed.

4. The autism diagnosis apparatus based on magnetic resonance supergraph contrast learning according to any one of claims 1-3, further comprising: Using the AAL1 atlas, the brain is divided into 116 regions; the average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels in the region; Subsequently, the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the Pearson correlation coefficient (PCC) between the pair of average time series is used to calculate the full functional connectivity (FC) matrix; the brain of each subject is represented as a supergraph using the AAL116 atlas.

5. The autism diagnosis device based on magnetic resonance supergraph contrastive learning according to claim 4, further comprising constructing a brain supergraph: The brain supergraph is constructed from the A116 atlas, which divides the fMRI image into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signal of all voxels in the region; then, the edge weight is calculated according to the Pearson correlation coefficient between the average time series of two different regions; The nodal features are derived from the amplitude of low-frequency fluctuations (ALFFs) in three frequency bands of the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and typical frequency band (0.01-0.08 Hz); ALFFs are defined as the total power in these low-frequency ranges, obtained from the Fourier transform of the mean time series; previous studies have demonstrated the effectiveness of using ALFFs as nodal features; the hypergraph constructed from this is denoted as wherein a set of symbolic nodes, representing a set of hyper-edges, a diagonal matrix representing edge weights, hypergraph with the association matrix is represented, and the rules for constructing this correlation matrix are described as follows: , wherein the degree of the node , define edge of degrees , hyperedge denoted as ; the diagonal matrices representing the node degrees and the hypergraph degrees are denoted by and representing; in In each, each element specifies the number of hyperedges that link to the corresponding node, while in In each element, each element represents the number of nodes associated with the corresponding hyperedge; the incidence matrix of the hypergraph is utilized It can be derived that the diagonal matrix and to encapsulate nodes and hyper-edges; given node features Then the node embedding is updated through the process of hyper-edge convolution; , wherein denotes node characteristics of a layer, are learnable parameters, is a nonlinear activation function.

6. The autism diagnosis device based on magnetic resonance supergraph contrastive learning according to claim 5, further comprising a masked supergraph encoder: Masked hypergraph encoder for node embedding learning and edge embedding: given fMRI data constructs hypergraph ; This hypergraph goes through a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer; each layer updates the node features through a message passing process; let a set of neighboring nodes of the node is denoted as The message passing process is then defined as follows: , wherein denotes a hyperedge convolution block that converts neighboring node features and edge weights into aggregated vectors; function is an MLP layer, which is a trainable function that maps the current node representation and the aggregated vector to a new representation; the message passing process is as follows: , The above process can be written in matrix form as follows: , wherein denotes normalized node features, matrix comprising corresponding row, and , , , are trainable parameters, , ; symbol denotes a batch normalization operation; the hypergraph module aims to learn a latent representation of the node features and then connecting them to form the embedded edge features 。 7. The autism diagnosis device based on magnetic resonance supergraph contrastive learning according to claim 6, further comprising a masked supergraph autoencoder: inputting the edge features into an MLP layer to generate parameters of a Bernoulli distribution and performing random edge dropout sampling on the parameters; the MLP layer includes one trainable weight and bias linear layer, followed by a activation function and the other has a weight and bias is a linear layer, and finally forms a sigmoid function that maps the output to the range (0, 1); this process is mathematically defined as: , In this way, the edge features are converted into scalars, representing the parameters of the Bernoulli distribution; Given parameters edge drop indicator for each edge Sampling was performed so that wherein indicates that the edge is discarded; matrix Forming a binary (0 or 1) Bernoulli mask B; To ensure gradient backpropagation, the sampling process needs to be reparameterized; The reparameterization trick of Luo et al. is used for this purpose; , wherein ， As a temperature parameter, governing the smoothness of the reparameterized sampling function; if falling in the interval inside, the expression inside the sigmoid function is positive if it lies in the interval If the value is inside, it is negative; therefore, the output is... The probability of convergence to 1 is ; as approaching 0, Become more and more similar to the binary indicator of sampling; applying a bernoulli mask to the correlation matrix by element-wise multiplication when at a time to a particular edge; thus, the input hypergraph is converted to an edge-deleted hypergraph, preserving the same number of nodes; the parameter matrix is generated from the hypergraph g = (V, E, W) The process is denoted as e, and the entire data augmentation process can be denoted as: 。 8. The autism diagnosis device based on magnetic resonance hypergraph contrast learning according to any one of claims 2-7, in the pre-training stage, the self-supervised adversarial contrast learning, the input hypergraph and its enhanced version are processed by the hypergraph encoder and the projection head to extract the hypergraph features, these inputs pass through a weight-shared edge encoder and node encoder block; In order to train the parameters in the feature extractor and the hypergraph enhancement process, a loss function is used, which forces the two feature vectors to approach if they come from the same hypergraph, and to move away if they come from different hypergraphs.

9. The autism diagnosis device based on magnetic resonance hypergraph contrast learning according to claim 8, the loss is defined by the infoMax principle, the goal is to maximize: , wherein represents a set of hypergraphs within a batch, Denote its cardinality; The similarity measure, denoted as sim, is defined as the cosine of the angle between the two input vectors: , maximizing This can be done by keeping all edges; to encourage more edges to be removed, a regularization term aims to encourage more edges to be removed and minimize all The average value of the value , Finally, the total objective function is defined as: , determined by gradient descent / ascent on respective parameters and The optimal value of.

10. The autism diagnosis device based on magnetic resonance hypergraph contrast learning according to any one of claims 1-9, hypergraph classification in the fine-tuning stage, in the fine-tuning stage, the pre-trained SSL-CLAH model is further optimized to enhance the feature representation, especially to identify important and abnormal regions related to ASD; Then the refined features from the SSL-CLAH block are input into the classification block; The final output is a probability-based decision indicating the likelihood of the subject having autism spectrum disorder (ASD) or healthy control (HC); This decision is intuitively represented in the form of probability, thereby helping the diagnosis process; The loss function used in the fine-tuning stage is the cross-entropy loss, defined as: , wherein a label representing a pth graph, representing a prediction of pth for a subject, N is the number of samples.

11. A method for autism diagnosis based on magnetic resonance hypergraph contrastive learning, comprising: Self-supervised learning stage, graph classification model (fine-tuning stage); The self-supervised stage includes hyperedge-aware contrastive learning and adversarial training (SSL-HCLA pre-training stage); The graph classification model includes a linear classifier, which classifies the extracted features, and the linear classifier distinguishes between ASD and HC subjects by analyzing brain connectivity patterns and functional relationships.

12. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 11, further comprising a data preprocessing stage, including several stages: estimating rs-fMRI reference images, correcting head motion, adjusting slice timing, and correcting sensitivity distortion.

13. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 12, the data preprocessing stage removes confounding factors by using inter-frame displacement, global signal, and average tissue signal as covariates, and after registering the fMRI volume to the standard MNI152 space, it is regressed.

14. The magnetic resonance hypergraph contrastive learning based autism diagnosis method of any one of claims 11-13, further comprising: The brain is segmented into 116 regions using the AAL1 atlas; the average time series (BOLD signal) of each region is obtained by averaging the time series of all voxels within that region; Subsequently, the amplitude of the amplitude of the low-frequency fluctuation (ALFF) node feature is derived from the average time series, and the Pearson correlation coefficient (PCC) between the pair of average time series is used to calculate the full functional connectivity (FC) matrix; the brain of each subject is represented as a hypergraph using the AAL116 atlas.

15. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 14, further comprising constructing a brain hypergraph: The brain hypergraph is constructed by the A116 atlas, which segments the fMRI image into multiple regions of interest (ROIs); each ROI is represented as a node, and the functional connectivity between any pair of ROIs constitutes the edge of the graph; the average time series within each ROI is calculated by averaging the BOLD signal of all voxels within the region; then, the edge weight is calculated according to the Pearson correlation coefficient between the average time series of two different regions; The nodal features are derived from the amplitude of low-frequency fluctuations (ALFFs) in three frequency bands of the BOLD signal: Slow-5 (0.01-0.027 Hz), Slow-4 (0.027-0.073 Hz), and typical frequency band (0.01-0.08 Hz); ALFFs are defined as the total power in these low-frequency ranges, obtained from the Fourier transform of the mean time series; previous studies have demonstrated the effectiveness of using ALFFs as nodal features; the hypergraph constructed from this is denoted as wherein a set of symbolic nodes, representing a set of hyper-edges, a diagonal matrix representing edge weights, hypergraph with the correlation matrix , the rule for constructing this correlation matrix is described as follows: , wherein the degree of the node , define edge of degrees , hyperedge denoted as ; the diagonal matrices representing the node degrees and the hypergraph degrees are denoted by and denotes; in In each, each element specifies the number of hyperedges that link to the corresponding node, while in In each of the above, each element represents the number of nodes associated with the corresponding hyperedge; the incidence matrix of the hypergraph is utilized It can be derived that the diagonal matrix and to encapsulate nodes and hyper-edges; given node features Then the node embedding is updated through the process of hyperedge convolution; , wherein denotes node characteristics of a layer, are learnable parameters, , represents a nonlinear activation function.

16. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 15, further comprising a mask hypergraph encoder: Masked hypergraph encoder for node embedding learning and edge embedding: given fMRI data constructs hypergraph ; This hypergraph goes through a series of transformations, including a hyperedge convolution block, feature concatenation, and an MLP layer; each layer updates the node features through a message passing process; let a set of neighboring nodes of the node is denoted as The message passing process is then defined as follows: ， wherein denotes a hyperedge convolution block that converts neighboring node features and edge weights into aggregated vectors; function is an MLP layer, which is a trainable function that maps the current node representation and the aggregated vector to a new representation; the message passing process is as follows: ， The above process can be written in matrix form as follows: ， wherein denotes normalized node features, matrix comprising corresponding row, and , , , are trainable parameters, , ; symbol denotes a batch normalization operation; the hypergraph module aims to learn a latent representation of the node features and then connecting them to form the embedded edge features 。 17. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 16, further comprising a mask hypergraph autoencoder: inputting the edge features into an MLP layer to generate parameters of a Bernoulli distribution and performing random edge dropout sampling on the parameters; the MLP layer includes one trainable weight and bias linear layer, followed by a activation function and the other has a weight and bias is a linear layer, and finally forms a sigmoid function that maps the output to the range (0, 1); this process is mathematically defined as: ， In this way, the edge feature is converted into a scalar, representing the parameter of the Bernoulli distribution; Given parameters Edge drop indicator for each edge Sampling was performed so that wherein indicates that the edge is discarded; matrix A binary (0 or 1) Bernoulli mask B is formed; in order to ensure gradient backpropagation, the sampling process needs to be reparameterized; the reparameterization trick used by Luo et al. is used for this purpose; ， wherein ， As a temperature parameter, governing the smoothness of the reparameterized sampling function; if Fall in interval inside, the expression inside the sigmoid function is positive if it lies in the interval If the value is inside, it is negative; therefore, the output is... The probability of convergence to 1 is ; as approaching 0, becomes more and more similar to the sampled binary indicator; applying a bernoulli mask to the correlation matrix by element-wise multiplication when at a time to a particular edge; thus, the input hypergraph is transformed into an edge-deleted hypergraph, preserving the same number of nodes; the parameter matrix is generated from the hypergraph g = (V, E, W) The process of is represented as e, and the entire data augmentation process can be represented as: 。 18. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to any one of claims 12-17, in the pre-training stage, the self-supervised adversarial contrast learning, the input hypergraph and its augmented version are processed by the hypergraph encoder and the projection head to extract the hypergraph features, these inputs pass through a weight-shared edge encoder and node encoder block; In order to train the parameters in the feature extractor and the hypergraph augmentation process, a loss function is used, which forces the two feature vectors to approach if they come from the same hypergraph, and to move away if they come from different hypergraphs.

19. The autism diagnosis method based on magnetic resonance hypergraph contrast learning according to claim 8, the loss is defined by the infoMax principle, the goal is to maximize: ， wherein represents a set of hypergraphs within a batch, denotes its cardinality; a similarity measure, denoted as sim, is defined as the cosine of the angle between two input vectors: ， maximizing This can be done by keeping all edges; to encourage more edges to be removed, a regularization term aims to encourage more edges to be removed and minimize all the average of the values ， Finally, the total objective function is defined as: ， determined by gradient descent / ascent on respective parameters and the optimal value of.

20. The method of autism diagnosis based on magnetic resonance hypergraph contrast learning according to any one of claims 11-19, the hypergraph classification in the fine-tuning stage, in the fine-tuning stage, the pre-trained SSL-CLAH model is further optimized to enhance the feature representation, especially to identify important and abnormal regions related to ASD; then the refined features from the SSL-CLAH block are input into the classification block; the final output is a probability-based decision indicating the likelihood of the subject having autism spectrum disorder (ASD) or healthy control (HC); this decision is intuitively represented in the form of probability, thereby facilitating the diagnosis process; the loss function used in the fine-tuning stage is cross-entropy loss, defined as: ， wherein a label representing a pth graph, representing a prediction of pth for a subject, N is the number of samples.

21. A storage medium for autism diagnosis based on magnetic resonance hypergraph contrast learning, having stored thereon a computer program (instructions) for implementing the method of autism diagnosis based on magnetic resonance hypergraph contrast learning according to any one of claims 11-20.

Images